Power transmission channel forest fire risk early warning method and system based on spatio-temporal feature fusion

Abstract

The application discloses a power transmission channel mountain fire risk early warning method and system based on space-time feature fusion, and relates to the technical field of risk early warning; the application generates a static flammability background map by constructing a power transmission channel risk grid and fusing historical multi-source data; by collecting real-time visual and micro-meteorological data, after space-time alignment processing, visual risk features and environmental flammability trend features are respectively extracted by using a double-branch space-time feature extraction network; the static background and dynamic features are input into a space-time fusion module to obtain a comprehensive mountain fire risk index; when the index exceeds a threshold, a mountain fire spread deduction model combined with a power transmission corridor effect correction is triggered to predict fire spread and evaluate line trip probability, and a graded early warning and emergency strategy is issued; the application overcomes the defects of serious data island and inaccurate early warning, realizes a leap from simple fire point monitoring to dynamic risk situation prediction, and significantly improves the accuracy of mountain fire early warning and the intelligent level of power grid prevention and control.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of risk early warning, and in particular relates to a power transmission channel forest fire risk early warning method and system based on spatio-temporal feature fusion. BACKGROUND

[0002] With the continuous expansion of the power grid, long-distance and large-capacity transmission lines are increasing, and their corridors often need to pass through forest, hilly and other vegetation dense areas. Forest fires have become one of the main natural disasters threatening the safe and stable operation of transmission lines, which not only may cause power grid failure trip, but also may even lead to tower collapse and other serious accidents, posing a serious threat to power grid safety and public power supply.

[0003] The present application is mainly applied to the prevention and control of forest fires in high-voltage and ultra-high-voltage power transmission channels. These channels are located in the wild and have complex terrain, and traditional manual inspection is inefficient and difficult to find fire in time. At present, although satellite remote sensing, video monitoring and micro-meteorological stations and other monitoring means have been widely used, how to effectively integrate these heterogeneous data to realize early and accurate early warning and situation deduction of forest fire risk is the core demand in the current transmission inspection field.

[0004] The existing forest fire monitoring and early warning methods mostly rely on a single data source or simple threshold judgment, which has obvious limitations; for example, satellite remote sensing has a long update cycle and is easily blocked by clouds; video monitoring relies on manual interpretation and is prone to missed reports and cannot quantify risks; micro-meteorological data can only reflect the environmental background and cannot directly perceive fire; more importantly, these technical means form a "data island", lack effective spatio-temporal correlation and fusion mechanism, resulting in high false alarm rate and weak pertinence of early warning, and it is difficult to dynamically evaluate the trend of fire spread and its specific harm to transmission lines, making it difficult to support decision-making upgrade from passive response to active early warning. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the problems in the related art, the present application provides a power transmission channel forest fire risk early warning method and system based on spatio-temporal feature fusion to overcome the above technical problems existing in the prior art.

[0007] (II) Technical solutions

[0008] To solve the above technical problems, the present application is realized by the following technical solutions:

[0009] In a first aspect, the present application provides a power transmission channel forest fire risk early warning method based on spatio-temporal feature fusion, comprising the following steps:

[0010] S1, divide the power transmission channel into a plurality of risk grids based on geographic information of the power transmission channel, and calculate a static flammability background of each risk grid in combination with historical multi-source data to obtain a static flammability background atlas;

[0011] S2, synchronously collect real-time visual data and real-time micro-meteorological data of the power transmission channel by using a plurality of sensor terminals deployed along the power transmission channel, and perform spatio-temporal alignment processing to obtain aligned real-time multi-source monitoring data;

[0012] S3, construct a double-branch spatio-temporal feature extraction network; and extract visual risk features and environmental flammability trend features from the aligned real-time multi-source monitoring data by using the double-branch spatio-temporal feature extraction network to obtain a visual risk confidence vector and an environmental flammability trend vector;

[0013] S4, input the static flammability background in the static flammability background atlas, the visual risk confidence vector and the environmental flammability trend vector into a spatio-temporal fusion module, perform feature fusion by using a spatio-temporal attention mechanism, and calculate a comprehensive forest fire risk index;

[0014] S5, when the comprehensive forest fire risk index exceeds a preset threshold, trigger a forest fire spread deduction model to predict a forest fire spread trend and a harm degree to the power transmission line, and issue a graded early warning and generate an emergency strategy according to the harm degree;

[0015] Preferably, the S1 comprises the following steps:

[0016] S11, establish a power transmission channel coordinate system based on a GIS geographic information system, extend a power transmission line corridor center line to both sides by a preset width, divide the power transmission channel into a plurality of continuous spatial grid units as risk grids, and take the power transmission line corridor center line as a reference;

[0017] S12, collect historical fire data, historical meteorological data, topographic and geomorphic data and vegetation remote sensing data of the power transmission channel;

[0018] S13, calculate historical combustible load and historical combustible moisture content of each risk grid based on the vegetation remote sensing data; and construct a vegetation fine feature atlas of the power transmission channel in combination with the topographic and geomorphic data and the vegetation type;

[0019] S14, train a basic fire risk classifier by using a random forest algorithm, taking the vegetation fine feature atlas and the historical meteorological data as inputs and taking a fire occurrence probability in the historical fire data as a label; and calculate a static flammability background level of each risk grid by using the trained basic fire risk classifier, and input the static flammability background level into the vegetation fine feature atlas to form a static flammability background atlas;

[0020] Preferably, the S13 comprises the following steps:

[0021] S131, acquire historical multi-spectral satellite remote sensing images of the power transmission channel area, extract reflectivity data of near-infrared and short-wave infrared bands; collect ground temperature data;

[0022] S132, calculate the normalized vegetation index based on the reflectivity data of the near-infrared and short-wave infrared bands;

[0023] S133, based on the normalized vegetation index and the ground temperature data, construct a vegetation water supply index formula; establish a regression model between the vegetation water supply index and the measured fuel moisture content; establish a regression model between the normalized vegetation index and the fuel load;

[0024] S134, use the above regression models to calculate the historical fuel moisture content and fuel load of each risk grid; combined with the vegetation type coefficient, the final historical fuel moisture content distribution and historical fuel load distribution of each risk grid are obtained;

[0025] S135, input the final historical fuel moisture content distribution, historical fuel load distribution, topographic data and vegetation type data into each corresponding risk grid to obtain a vegetation fine feature map;

[0026] Preferably, the S2 comprises the following steps:

[0027] S21, deploying a multi-sensor terminal on the tower of the power transmission channel; the multi-sensor terminal includes a visible light camera, an infrared thermal imager, and a micro-meteorological station;

[0028] S22, collecting real-time video stream data of the power transmission channel using the visible light camera to obtain real-time video stream data; collecting real-time infrared thermal image data using the infrared thermal imager to obtain real-time infrared thermal image data; collecting real-time wind speed, wind direction, temperature, humidity, and air pressure data using the micro-meteorological station to obtain real-time micro-meteorological data; the real-time infrared thermal image data and real-time video stream data constitute real-time visual data;

[0029] S23, performing spatio-temporal alignment processing on the collected real-time video stream data and real-time infrared thermal image data; mapping video stream data and micro-meteorological data of different sampling frequencies to the same time window; using a perspective transformation matrix, mapping two-dimensional image pixel coordinates to a three-dimensional geographic coordinate system to realize spatial alignment of visual data and risk grids;

[0030] S24, taking the spatially aligned real-time video stream data, real-time infrared thermal image data, and real-time micro-meteorological data as aligned real-time multi-source monitoring data;

[0031] Preferably, the S3 comprises the following steps:

[0032] S31. Construct a dual-branch spatiotemporal feature extraction network; the dual-branch spatiotemporal feature extraction network includes a visual branch network and an environment branch network; the visual branch network adopts a YOLOv5 lightweight convolutional neural network, and adds an infrared channel at the input end of the YOLOv5 lightweight convolutional neural network, superimposing the visible light RGB three channels with the infrared single channel into a 4-channel input; the environment branch network adopts a long short-term memory network;

[0033] S32. Input the aligned real-time video stream data and real-time infrared thermal image data into the visual branch network, extract visual risk features, and obtain a visual risk confidence vector; the visual risk features include visible light features and infrared features.

[0034] S33. The aligned micro-meteorological data is input into the environmental branch network. The LSTM network uses its gating mechanism to capture the sudden increase trend of wind speed, the sudden decrease trend of humidity and the continuous increase trend of temperature in the time series, and obtains the environmental flammability trend vector through the fully connected layer mapping.

[0035] Preferably, step S4 includes the following steps:

[0036] S41. Construct a spatiotemporal fusion module; the spatiotemporal fusion module includes a spatial attention unit and a temporal attention unit;

[0037] S42. Input the static flammability baseline, visual risk confidence vector, and environmental flammability trend vector into the spatiotemporal fusion module; calculate the spatial weight matrix using the spatial attention unit and calculate the temporal weight coefficient using the temporal attention unit;

[0038] S43. Based on the spatial weight matrix and time weight coefficient, the static flammability background, visual risk confidence vector and environmental flammability trend vector are weighted and fused to calculate the comprehensive wildfire risk index for each risk grid.

[0039] Preferably, step S42 includes the following steps:

[0040] S421. Calculate the ratio of the number of grids located in the current real-time wind direction or at an altitude higher than the current risk grid in all neighboring grids of any risk grid to the total number of neighboring grids of any risk grid, and obtain the wind direction slope weight.

[0041] S422, The time attention unit is a fully connected neural network; collect historical wildfire case data; the historical wildfire case data includes the time period characteristics, visual risk confidence, environmental flammability trend value and static flammability background of each case;

[0042] Based on the actual fire spread of each case in the historical wildfire case data, a historical time weighting coefficient is set; the historical time weighting coefficient is set as the label of the historical wildfire case data to obtain labeled historical wildfire case data.

[0043] The fully connected neural network was trained using labeled historical wildfire case data to obtain the final fully connected neural network.

[0044] The real-time time period features, visual risk confidence, environmental flammability trend value and static flammability background are input into the final fully connected neural network to obtain the real-time time weight coefficients.

[0045] Preferably, step S5 includes the following steps:

[0046] S51. Set wildfire risk thresholds; monitor the comprehensive wildfire risk index of each risk grid in real time. When the comprehensive wildfire risk index exceeds the wildfire risk threshold, the grid is determined to be the ignition point, triggering the wildfire spread simulation model.

[0047] S52. Construct a wildfire spread simulation model based on the Rothermel model; introduce a power transmission corridor effect correction coefficient to correct the wind speed parameter in the Rothermel model, and obtain the corrected wildfire spread simulation model.

[0048] S53. Using the modified wildfire spread simulation model, with the ignition point as the center, combined with real-time micro-meteorological data and topographic data, predict the fire spread path, flame height and smoke concentration distribution within a preset time period in the future.

[0049] S54. Establish a tripping probability assessment model for transmission lines; calculate the tripping probability of air gap breakdown in transmission lines based on the predicted flame height, smoke concentration, conductor-to-ground distance, and insulator string length.

[0050] S55. Construct a wildfire risk level; based on the wildfire risk level, combined with the comprehensive wildfire risk index and the power outage probability, generate corresponding emergency strategies;

[0051] Preferably, S54 includes the following steps:

[0052] S541. Establish a formula for calculating the air gap insulation strength under flame and smoke environments; based on the flame height and smoke concentration distribution, calculate the predicted air breakdown voltage at future moments using the air gap insulation strength calculation formula.

[0053] S542. Obtain the operating voltage of the transmission line; calculate the tripping probability at each future moment based on the operating voltage of the transmission line and the predicted air breakdown voltage at future moments.

[0054] Secondly, the present invention also provides a power transmission channel wildfire risk early warning system based on spatiotemporal feature fusion, used to implement the above-mentioned power transmission channel wildfire risk early warning method based on spatiotemporal feature fusion, the system comprising:

[0055] Multidimensional data acquisition and static baseline construction module: Divide the power transmission channel into regular risk grids and collect historical multi-source data; calculate and generate the static flammability baseline level for each risk grid, and construct a static flammability baseline map;

[0056] Real-time multi-source monitoring and data preprocessing module: By deploying multi-sensor terminals on transmission towers, real-time visual information and micro-meteorological data of the transmission channel are collected synchronously; by time window alignment and perspective transformation-based spatial coordinate mapping technology, aligned real-time multi-source monitoring data is generated.

[0057] Dynamic risk feature extraction module: Based on aligned real-time multi-source monitoring data, the visual branch network outputs a visual risk confidence vector; the environmental branch network outputs an environmental flammability trend vector by capturing trend features that predict the deterioration of environmental flammability.

[0058] The spatiotemporal fusion and risk calculation module: The static flammability baseline, visual risk confidence vector, and environmental flammability trend vector from the static flammability baseline map are input into the spatiotemporal fusion module. The spatial attention unit calculates the spatial weight matrix based on real-time wind direction, slope, and combustible material distribution to quantify the spatial spread tendency of the fire. The temporal attention unit generates temporal fusion weight coefficients based on time periodicity and real-time risk evidence. Through a weighted fusion algorithm, the comprehensive wildfire risk index of each risk grid is output.

[0059] Wildfire spread simulation and graded early warning module: When the comprehensive wildfire risk index exceeds the threshold, the wildfire spread simulation model is triggered to accurately predict the fire path, flame height and smoke concentration; a tripping probability assessment model is constructed in combination with transmission line parameters and the tripping probability is calculated; emergency strategies are generated based on the comprehensive wildfire risk index and the tripping probability.

[0060] (III) Beneficial Effects

[0061] The present invention has the following beneficial effects:

[0062] (a) Beneficial effects of the present invention

[0063] This invention achieves multi-dimensional and refined perception and early warning of wildfire risks by constructing a risk assessment system that combines static baseline data with dynamic disturbances. By solidifying historical patterns into a static flammability baseline map and deeply integrating it with real-time perceived visual and environmental dynamic features, it overcomes the limitations of existing technologies that rely on a single data source. This dynamic and static combined assessment mechanism can identify potential fire hazards earlier, significantly improve the foresight and accuracy of early warnings, and realize a fundamental shift from passive response to proactive prediction.

[0064] This invention utilizes a spatiotemporal attention mechanism and a deep learning model to effectively solve the "island" problem of multi-source heterogeneous data and realize intelligent risk assessment. The dual-branch network performs specialized feature extraction for the characteristics of image and time-series data, ensuring the effective capture of fireworks features and meteorological trend features. The spatiotemporal fusion module dynamically calculates weights and adaptively adjusts the contribution of different evidence in different spatiotemporal contexts, simulating expert decision-making logic, significantly reducing false alarms and false negatives, and making the risk assessment results more robust and interpretable.

[0065] This invention combines a wildfire physical spread model with a power grid safety assessment model, achieving closed-loop management from fire monitoring to power grid risk pre-control. By introducing a transmission corridor effect to correct the spread model and constructing a tripping probability assessment model based on effective insulation distance, it can accurately predict the specific degree of harm and remaining safe time of wildfires to transmission lines. The generation of tiered early warning and emergency strategies provides operation and maintenance personnel with intuitive and quantitative decision support, effectively ensuring the safe and stable operation of the power grid and improving the intelligence level and emergency response efficiency of wildfire prevention and control.

[0066] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0067] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0068] Figure 1 This is a flowchart illustrating the method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion according to the present invention.

[0069] Figure 2 This is a schematic diagram of the module of the power transmission channel wildfire risk early warning system based on spatiotemporal feature fusion of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0071] To resolve the above issues, please refer to [link / reference]. Figure 1 This invention discloses a method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion, comprising the following steps:

[0072] S1. Based on the geographic information of the power transmission channel, the power transmission channel is divided into several risk grids, and combined with historical multi-source data, the static flammability background of each risk grid is calculated to obtain a static flammability background map.

[0073] S2. Using multi-sensor terminals deployed along the power transmission channel, real-time visual data and real-time micro-meteorological data of the power transmission channel are collected synchronously and spatiotemporally aligned to obtain aligned real-time multi-source monitoring data.

[0074] S3. Construct a dual-branch spatiotemporal feature extraction network; use the dual-branch spatiotemporal feature extraction network to extract visual risk features and environmental flammability trend features from the aligned real-time multi-source monitoring data, respectively, to obtain the visual risk confidence vector and the environmental flammability trend vector;

[0075] S4. Input the static flammability baseline, visual risk confidence vector and environmental flammability trend vector from the static flammability baseline map into the spatiotemporal fusion module, use the spatiotemporal attention mechanism to perform feature fusion, and calculate the comprehensive wildfire risk index.

[0076] S5. When the comprehensive wildfire risk index exceeds the preset threshold, the wildfire spread simulation model is triggered to predict the wildfire spread trend and the degree of harm to power transmission lines, and graded early warnings are issued and emergency strategies are generated according to the degree of harm.

[0077] The above embodiments address the problems of severe data silos, weak spatiotemporal correlation, and lack of specificity in existing power transmission channel wildfire monitoring, which leads to inaccurate early warnings. By constructing a risk assessment system that combines static baseline data with dynamic disturbances, and by deeply fusing image visual features and meteorological environmental features using a spatiotemporal attention network, and combining the unique corridor effect of power transmission channels for spread extrapolation, the embodiments achieve a leap from simple fire point monitoring to dynamic prediction of risk situations.

[0078] Step S1 above includes the following steps:

[0079] S11. Establish a coordinate system for the power transmission channel based on the GIS geographic information system. Using the center line of the power transmission line corridor as the reference, extend the preset width to both sides to divide the power transmission channel into several continuous spatial grid units as risk grids.

[0080] S12. Collect historical fire data, historical meteorological data, topographic data, and vegetation remote sensing data of the power transmission channel;

[0081] S13. Based on vegetation remote sensing data, invert and calculate the historical combustible load and historical combustible moisture content of each risk grid; combine topographic data and vegetation type to construct a refined vegetation feature map of the power transmission channel.

[0082] S14. Using the random forest algorithm, with the vegetation fine feature map and historical meteorological data as input, and the fire occurrence probability in the historical fire data as the label, train a basic fire risk classifier; use the trained basic fire risk classifier to calculate the static flammability baseline level of each risk grid, and input the static flammability baseline level into the vegetation fine feature map to form a static flammability baseline map.

[0083] In specific implementation, step S11 is as follows: Select a certain ultra-high voltage direct current transmission channel as the implementation object; import the tower coordinate data of the line using ArcGIS software; extend 2 kilometers to the left and right sides of the line centerline (covering the safe distance where wildfires may spread to the line), forming a strip monitoring area with a width of 4 kilometers; divide the area into 500m×500m square grid units; assign a unique ID number to each grid and record the latitude and longitude coordinates of its center point and the tower interval to which it belongs (e.g., between towers 105 and 106).

[0084] In specific implementation, step S12 is as follows: The historical fire data comes from the power grid company's fault trip records and the forestry department's fire statistics yearbook, including the ignition time (accurate to the minute), ignition location coordinates, burned area, and cause of fire (lightning strike, sacrificial ceremony, agricultural fire, etc.) of all wildfire events that occurred in the area in the past 5 years.

[0085] The historical meteorological data are derived from the historical records of the micro-meteorological stations built by the meteorological bureau and the power grid, including the daily average temperature, relative humidity, maximum wind speed, wind direction, and 24-hour rainfall for the same period in history.

[0086] The topographic data comes from a high-precision digital elevation model (DEM), and the average elevation, slope (degrees), and aspect (0-360 degrees) of each grid are extracted.

[0087] The vegetation remote sensing data are derived from MODIS or Landsat satellite imagery, including Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), and Enhanced Vegetation Index (EVI).

[0088] In specific implementation, step S14 above is as follows: Construct a training sample set, with input features including: vegetation type, combustible load, water content, slope, aspect, historical average temperature for the same period, and historical average rainfall for the same period; the label is whether a fire has occurred in the grid during the same period in history (0 or 1); train a classification model using the random forest algorithm; input the current static features of all grids in the transmission channel into the model, and output the fire occurrence probability of each grid; divide the probability into 5 levels (Level 1 low risk - Level 5 extremely high risk) as the static flammability baseline;

[0089] The above embodiments establish a refined gridded database, solidifying historical patterns and geographical environmental characteristics into a static baseline, providing a benchmark reference for subsequent dynamic risk assessment; while making full use of the value of historical data and quantifying the inherent risk level of the region, they also achieve a refined characterization of the differences in flammability along the power transmission channel, effectively avoiding a one-size-fits-all approach to early warning.

[0090] Step S13 above includes the following steps:

[0091] S131. Acquire historical multispectral satellite remote sensing images of the power transmission channel area, extract reflectance data in the near-infrared and short-wave infrared bands; collect surface temperature data;

[0092] S132. Based on reflectance data from the near-infrared and short-wave infrared bands, calculate the Normalized Difference Vegetation Index (NDVI); the calculation formula is as follows:

[0093] NDVI = (NIR - R) / (NIR + R)

[0094] NIR and R represent the near-infrared and red light reflectance data, respectively, in the near-infrared and short-wave infrared bands. The larger the NDVI value, the denser the vegetation and the greater the potential combustible load.

[0095] S133. Based on the Normalized Difference Vegetation Index (NDVI) and surface temperature data, a vegetation water supply index formula is constructed; a regression model is established between the vegetation water supply index (VSVI) and the measured combustible moisture content (FMC); a regression model is established between the normalized vegetation index (NDVI) and the combustible load (FL); the vegetation water supply index formula is as follows:

[0096] VSWI = NDVI / LST;

[0097] The regression model formula between the vegetation water supply index (VSWI) and the measured combustible material moisture content (FMC) is as follows:

[0098] FMC= a ×VSWI+b, where a b are the first regression coefficients, and the regression coefficients are... a b. By setting up several sample plots on the ground, the fresh weight and dry weight of the vegetation were measured regularly, the actual moisture content (FMC) was calculated, and the value was fitted and confirmed with the VSWI value; in this embodiment, the fitted result was FMC = 0.85 × VSWI + 0.12.

[0099] The regression model for the Normalized Difference Vegetation Index (NDVI) and Combustible Material Load (FL) is as follows:

[0100] FL = c × NDVI d +e; where c, d, and e represent the second regression coefficients; c, d, and e are obtained by setting up several sample plots on the ground, periodically measuring the dry weight of vegetation per unit area to obtain the combustible load, and confirming it by nonlinear regression fitting with the synchronous NDVI value.

[0101] S134. Using the above regression model, calculate the historical combustible moisture content and combustible load for each risk grid. Combined with the vegetation type coefficient, correct the final historical combustible moisture content distribution and historical combustible load distribution for each risk grid. Because different vegetation types (coniferous forest, broadleaf forest, shrubland, grassland) have different water retention capacities, a vegetation type correction coefficient is introduced to obtain the final historical combustible moisture content (FMC). final =k veg1 ×FMC, Final Historical Combustible Material Loading (FMC) final =k veg2 ×FMC; Correction coefficient k for the first vegetation type veg1 The water retention capacity of different vegetation types is quantified and normalized. For example, under the same environmental conditions, standard samples from different vegetation types are given the same initial water volume and their water retention rates are measured after the same time period. The measurement results are then normalized. For instance, using grassland as the baseline (coefficient 1.0), under the same conditions, the final water retention rate of broadleaf forest samples is 1.1 times that of grassland and 1.22 times that of coniferous forest samples. Therefore, the correction coefficient for broadleaf forest is set to 1.1, and for coniferous forest to 0.9. The correction coefficient for the second vegetation type is k. veg1 This was obtained by normalizing the difference in unit biomass density.

[0102] S135. Input the final historical combustible moisture content distribution, historical combustible load distribution, topographic data and vegetation type data into each corresponding risk grid to obtain a refined vegetation feature map; the topographic data includes altitude, slope and aspect; the vegetation type data includes vegetation type, NDVI, combustible load and combustible moisture content.

[0103] S2. Using multi-sensor terminals deployed along the power transmission channel, real-time visual data and real-time micro-meteorological data of the power transmission channel are collected synchronously and spatiotemporally aligned to obtain aligned real-time multi-source monitoring data.

[0104] Step S2 above includes the following steps:

[0105] S21. Deploy multi-sensor terminals on the towers of the power transmission channel; the multi-sensor terminals include visible light cameras, infrared thermal imagers, and micro-weather stations;

[0106] S22. Real-time video stream data of the power transmission channel is acquired using a visible light camera; real-time infrared thermal image data is acquired using an infrared thermal imager; real-time wind speed, wind direction, temperature, humidity, and air pressure data are acquired using a micro-weather station; the real-time infrared thermal image data and the real-time video stream data constitute real-time visual data; the sampling frequency of the visible light and infrared data is 25 frames / second; the sampling frequency of the micro-weather data is 1 time / minute.

[0107] S23. Perform spatiotemporal alignment processing on the collected real-time video stream data and real-time infrared thermal image data; map video stream data and micro-meteorological data with different sampling frequencies to the same time window; use the perspective transformation matrix to map the pixel coordinates of the two-dimensional image to the three-dimensional geospatial coordinate system to achieve spatial alignment between visual data and risk grid.

[0108] S24. Use the spatially aligned real-time video stream data, real-time infrared thermal image data and real-time micro-meteorological data as aligned real-time multi-source monitoring data.

[0109] In specific implementation, step S21 is as follows: a multi-sensor terminal is deployed every 3 poles (approximately 1-1.5 kilometers) in key sections; the visible light camera is a high-definition zoom PTZ camera that supports 360-degree cruise; the infrared thermal imager is coaxially mounted with the visible light camera, with overlapping field of view; the micro-weather station is installed on the pole body and has lightning protection and electromagnetic interference protection capabilities.

[0110] In specific implementation, the time alignment in step S23 above is as follows: set a uniform time window (e.g., 1 minute); within this window, take the average value of the micro-meteorological data as the meteorological feature of that minute; take the key frames of the video stream (e.g., the 1st second and 30th second of each minute) as the visual feature input;

[0111] The spatial alignment is specifically as follows: Since the camera captures a two-dimensional image, while the risk grid is a three-dimensional geographic space; using pre-calibrated camera intrinsic parameters (focal length, principal point) and extrinsic parameters (installation height, pitch angle, azimuth angle), combined with the elevation model DEM, a perspective transformation matrix is ​​constructed. For each pixel in the image, its corresponding geographic coordinates are calculated using a monocular vision positioning algorithm, thereby determining which risk grid ID the smoke pixel found in the image falls into.

[0112] The above embodiments solve the problem of asynchronous heterogeneous data in time and space by using multi-sensor fusion acquisition and spatiotemporal alignment technology; while achieving the construction of a real-time monitoring dataset with a unified spatiotemporal benchmark, it also realizes the effective complementarity of visual intuitive information and meteorological environmental parameters, laying a data foundation for subsequent multimodal feature fusion.

[0113] S3. Construct a dual-branch spatiotemporal feature extraction network; use the dual-branch spatiotemporal feature extraction network to extract visual risk features and environmental flammability trend features from the aligned real-time multi-source monitoring data, respectively, to obtain the visual risk confidence vector and the environmental flammability trend vector;

[0114] Step S3 above includes the following steps:

[0115] S31. Construct a dual-branch spatiotemporal feature extraction network; the dual-branch spatiotemporal feature extraction network includes a visual branch network and an environment branch network; the visual branch network adopts a YOLOv5 lightweight convolutional neural network, and adds an infrared channel at the input end of the YOLOv5 lightweight convolutional neural network, superimposing the visible light RGB three channels with the infrared single channel into a 4-channel input; the environment branch network adopts a long short-term memory network;

[0116] S32. Input the aligned real-time video stream data and real-time infrared thermal image data into the visual branch network, extract visual risk features, and obtain a visual risk confidence vector; the visual risk features include visible light features and infrared features.

[0117] S33. The aligned micro-meteorological data is input into the environmental branch network. The LSTM network uses its gating mechanism (forget gate, input gate, output gate) to capture the sudden increase trend of wind speed, the sudden decrease trend of humidity and the continuous increase trend of temperature in the time series. The environmental flammability trend vector is obtained through mapping by the fully connected layer.

[0118] In specific implementation, step S32 is as follows: The aligned real-time video stream data and real-time infrared thermal image data are input into the visual branch network; the network first extracts low-level features through convolutional layers, learning the texture and color features of smoke and the shape and brightness features of flames from the visible light channel, while simultaneously identifying abnormal high-temperature area features from the infrared channel. The detection head outputs a comprehensive vector containing smoke confidence, open flame confidence, high temperature confidence, and the position coordinates of the target box; corresponding weight coefficients are set based on the hazard levels of smoke, open flame, and high temperature; the corresponding risk grid is located using the position coordinates of the target box; and the visual risk confidence vector V for each risk grid is calculated based on the weight coefficients, smoke confidence, open flame confidence, and high temperature confidence. score ;

[0119] In specific implementation, step S33 is as follows: Construct a time sliding window (e.g., a length of 60, representing the past 60 minutes), and extract real-time micro-meteorological data sequences (wind speed, humidity, temperature) from the past hour; input the micro-meteorological data sequences into the environmental branch network. The LSTM network utilizes its gating mechanism (forget gate, input gate, output gate) to capture the long-short dependencies in the time series, identifying whether wind speed increases sharply from low speed in a short period (instantaneous rate of change of wind speed), whether relative humidity continuously decreases (e.g., from 60% to 30%, indicating rapid drying of combustibles and increased flammability) (slope of continuous humidity decrease), and identifying abnormal increases in temperature (cumulative increase in temperature) to obtain environmental flammability trend characteristics; map the environmental flammability trend characteristics through a fully connected layer and normalize them using a Sigmoid activation function to obtain an environmental flammability trend vector; the mapping and normalization specifically results in the environmental flammability trend vector E. score =Sigmoid(W×h t +b); where W represents the weight matrix of the fully connected layer, b represents the bias term, and Sigmoid represents the activation function; ht represents the weighted sum of wind speed, humidity, and temperature changes in the environmental flammability trend characteristics. The specific weight values ​​can be dynamically set according to the application, and will not be elaborated further here.

[0120] The above embodiments employ a dual-branch network structure to process visual images and meteorological time-series data respectively, and design dedicated feature extractors for the characteristics of different modal data; while fully mining the texture, color, and temperature features in the images and the changing trend features in the meteorological data, it achieves comprehensive capture of the direct characterization (smoke and fire) and indirect causes (meteorological changes) of wildfires.

[0121] S4. Input the static flammability baseline, visual risk confidence vector and environmental flammability trend vector from the static flammability baseline map into the spatiotemporal fusion module, use the spatiotemporal attention mechanism to perform feature fusion, and calculate the comprehensive wildfire risk index.

[0122] Step S4 above includes the following steps:

[0123] S41. Construct a spatiotemporal fusion module; the spatiotemporal fusion module includes a spatial attention unit and a temporal attention unit;

[0124] S42. Input the static flammability baseline, visual risk confidence vector, and environmental flammability trend vector into the spatiotemporal fusion module; calculate the spatial weight matrix using the spatial attention unit and calculate the temporal weight coefficient using the temporal attention unit;

[0125] S43. Based on the spatial weight matrix and time weight coefficient, the static flammability baseline, visual risk confidence vector, and environmental flammability trend vector are weighted and fused to calculate the comprehensive wildfire risk index (WFRI) for each risk grid; the weighted fusion formula is as follows:

[0126] WFRI(i)=W sp (i)×(q1×V score (i)+q2×E score (i)+q3×B static (i)); where WFRI(i) is the comprehensive wildfire risk index of the i-th risk grid; V score (i) represents the visual risk confidence vector of the i-th risk grid, E score (i) represents the environmental flammability trend vector of the i-th risk grid;

[0127] The above embodiments introduce a spatiotemporal attention mechanism to dynamically adjust the weights of different data sources under different spatiotemporal backgrounds; achieving the effect of simulating the judgment logic of human experts (such as paying more attention to weather on dry and windy days, and paying more attention to vision when seeing open flames), realizing the adaptability and high robustness of risk assessment, and solving the problem of false alarms and false negatives caused by single fixed weight fusion.

[0128] Step S42 above includes the following steps:

[0129] S421. For any risk grid i, the spatial weight W due to wind direction and slope. ds (i) Determined by the number of neighboring grids of any risk grid i in the upwind and upslope directions; Calculate the number N of all neighboring grids of any risk grid that are located upwind of the current real-time wind direction (within ±45 degrees) or at an altitude higher than it (slope > 0). risk The total number of neighborhood grids N of any risk gridtotal The ratio of wind direction to slope weight is used to obtain the wind direction slope weight; the calculation formula is as follows:

[0130] W ds (i)=1.0+N risk / N total ;W ds (i) represents the wind direction and slope weight of the i-th risk grid; and is based on the static flammability background level B of any risk grid i. static (i) The combustible weight W is directly assigned to any risk grid using the weighting formula. fu (i); The weighting formula is as follows:

[0131] W fu (i) = 0.8 + 0.2 × B static (i);

[0132] The spatial weight is obtained by combining the wind direction and slope weight with the combustible material weight; the calculation formula is as follows:

[0133] W sp (i)=W ds (i)×W fu (i); Wsp(i) represents the spatial weight of the i-th risk grid;

[0134] S422, The time attention unit is a fully connected neural network; historical wildfire case data is collected; the historical wildfire case data includes the time period characteristics and visual risk confidence V of each case. score Environmental flammability trend value E score And static flammability background; the time period characteristics are obtained through the following steps: mapping a year to a periodic function, extracting the sine and cosine values ​​of the date in the annual cycle (sin(2πdoy / 365), cos(2πdoy / 365)), and the number of hours in a day; where: doy represents the day of year, that is, which day of the year;

[0135] Based on the actual fire spread of each case in the historical wildfire case data, a historical time weighting coefficient is set; the historical time weighting coefficient is set as the label of the historical wildfire case data to obtain labeled historical wildfire case data.

[0136] The fully connected neural network was trained using labeled historical wildfire case data to obtain the final fully connected neural network.

[0137] The real-time time period features, visual risk confidence, environmental flammability trend value and static flammability background are input into the final fully connected neural network to obtain the real-time time weight coefficients (q1, q2, q3).

[0138] S5. Wildfire spread simulation and graded early warning based on transmission corridor effect; when the comprehensive wildfire risk index exceeds the preset threshold, the wildfire spread simulation model is triggered to predict the wildfire spread trend and the degree of damage to transmission lines, and graded early warnings and emergency strategies are issued according to the degree of damage.

[0139] Step S5 above includes the following steps:

[0140] S51. Set a wildfire risk threshold; monitor the comprehensive wildfire risk index (WFRI) of each risk grid in real time. When the comprehensive wildfire risk index (WFRI) exceeds the wildfire risk threshold (e.g., 0.75), the grid is determined to be an ignition point, triggering the wildfire spread simulation model. The wildfire risk threshold is set by statistically analyzing the distribution of the comprehensive wildfire risk index (WFRI) output by the early warning model before the occurrence of historical wildfire cases, and determining the optimal classification performance point based on the ROC curve.

[0141] S52. Construct a wildfire spread simulation model based on the Rothermel model; introduce a power transmission corridor effect correction coefficient to correct the wind speed parameter in the Rothermel model, and obtain the corrected wildfire spread simulation model.

[0142] S53. Using the modified wildfire spread simulation model, with the ignition point as the center, combined with real-time micro-meteorological data and topographic data, predict the fire spread path, flame height and smoke concentration distribution within a preset time period in the future.

[0143] S54. Establish a transmission line tripping probability assessment model; based on the predicted flame height, smoke concentration, and the conductor-to-ground distance H of the transmission line. gap Insulator string length L il Calculate the tripping probability of air gap breakdown in transmission lines;

[0144] S55. Construct a wildfire risk level; based on the wildfire risk level, combine the comprehensive wildfire risk index (WFRI) and the tripping probability (P). trip Generate corresponding emergency strategies;

[0145] In specific implementation, step S53 above is as follows: the modified wildfire spread simulation model is simulated using the cellular automata method; each risk grid is treated as a cell, and the spread rate is modified according to the modified wildfire spread simulation model. R Based on the direction and direction of the fire, predict which neighboring cells the fire will ignite in the next moment, and simultaneously calculate the flame height H of each burning cell. flame and smoke concentration C smoke The formula for calculating the flame height is as follows:

[0146] H flame =0.45×(I / 1000)0.46 Where I represents the fire intensity (kW / m); I = H × w m ×R; where H represents the lower calorific value (kJ / kg) of the combustible material of the vegetation within each risk grid, for example, for common herbs and shrubs, the value is between 18000-20000 kJ / kg; w m The combustible load of the cell (kg / m²) is obtained from step S134;

[0147] The formula for calculating the smoke concentration distribution is as follows:

[0148] C smoke =η×w m ×A m Where η represents the flue gas emission factor of the vegetation in each risk grid, and A m The burning area (m²) of this cell is the area of ​​a risk grid in a cellular automaton.

[0149] In specific implementation, step S55 above is as follows:

[0150] Blue alert (attention level): 0.75 ≥ WFRI > 0.6, P trip <10%; Strategy: Notify the line maintenance personnel and continuously monitor the situation from the back end;

[0151] Yellow Alert (Level 1): 0.85 ≥ WFRI > 0.75, 10% ≤ P trip <30%; Strategy: Line guards must arrive at the scene within 30 minutes to verify, and drones will be used for reconnaissance;

[0152] Orange Alert (Hazard Level): 0.9 ≥ WFRI > 0.85, 30% ≤ P trip <50%; Strategy: The dispatch center limits the load on this line, prepares to start the backup power supply, and the fire department is dispatched.

[0153] Red Alert (Emergency Level): WFRI > 0.9, 50% ≤ P trip <80%; Strategy: Immediately apply for line shutdown to avoid danger and prevent power grid oscillation caused by short circuit, and at the same time activate the automatic fire extinguishing device (if it is installed on the tower).

[0154] The above embodiments introduce a "corridor effect" correction model for special terrain of transmission channels and combine it with electrical insulation strength theory to assess the probability of tripping; thus, they achieve the transformation from simply predicting physical parameters of fire to predicting the safety risks of power grid operation, realize refined early warning for the characteristics of transmission lines, and provide direct decision-making basis for dispatching and operation personnel.

[0155] Step S54 above includes the following steps:

[0156] S541. Establish a formula for calculating the air gap insulation strength under flame and smoke environments; based on the flame height and smoke concentration distribution, calculate the predicted air breakdown voltage at future moments using the air gap insulation strength calculation formula; the air gap insulation strength calculation formula is as follows:

[0157] U bd (t)=k1×exp(-k2×Csmoke(t))×(min(H gap ,k3×L il )-Hflame(t));

[0158] Among them, U bd (t), C smoke (t) and H flame (t) represents the air breakdown voltage, smoke concentration, and flame height at future time t, respectively; k1 is the reference air breakdown field strength, characterizing the insulation strength of the air gap under standard atmospheric conditions; k2 is the smoke attenuation coefficient, characterizing the rate attenuation of air insulation strength per unit smoke concentration; k3 is the insulation strength conversion factor, characterizing the equivalent air insulation distance per unit length of insulator string in a wildfire environment, its value is determined by comparing the power frequency breakdown test between the insulator string and the air gap; min(H gap ,k3×L il () represents the effective insulation distance, and its value is determined by considering the parallel risks of air gap breakdown and insulator string flashover under wildfire conditions, taking the critical distance corresponding to the path with lower insulation strength between the two; for example, for a certain line H gap =15m, L il =5m, k3=0.7; then the effective insulation distance =min(15,0.7×5)=min(15,3.5)=3.5m; this result shows that in this environment, the flashover of the insulator string is the weakest link, and the effective insulation distance of the system is determined by it. 3.5 meters should be used in the calculation.

[0159] S542. Obtain the operating voltage U of the transmission line. operate According to the operating voltage U of the transmission line operate Based on the predicted air breakdown voltage at future times, the tripping probability at each future time is calculated; the calculation process is as follows:

[0160] When U operate >U bd If (t), breakdown will inevitably occur, and the tripping probability P trip =1; when U operate ≤U bd (t), due to the randomness of atmospheric conditions, there is still a certain probability of tripping; the tripping probability P is calculated using an exponential distribution model. trip The exponential distribution model calculates the tripping probability; the calculation formula for the exponential distribution model is as follows:

[0161] P trip (t)=A p ×exp(B p ×(U operate / U bd (t))); where A and B are probability distribution parameters; by fitting historical wildfire tripping cases, under the known line operating voltage U operate and the breakdown voltage U derived from the flame and smoke conditions bd The probability distribution of tripping events is determined; in this embodiment, the values ​​are A... p =0.15, B p =3.5;

[0162] S543. Based on the tripping probability at each future moment, generate a tripping risk time series curve and determine the remaining safe time; for example, if it is predicted that the flame height will reach 5 meters in 10 minutes, then P trip If the safety margin rises to 80%, the "remaining safe time" is 10 minutes.

[0163] For further details, please refer to Figure 2 A wildfire risk early warning system for power transmission channels based on spatiotemporal feature fusion is used to implement the aforementioned wildfire risk early warning method for power transmission channels based on spatiotemporal feature fusion, including:

[0164] Multidimensional data acquisition and static baseline construction module: This module is used to construct a static basic database for wildfire risk assessment of power transmission corridors. Based on GIS, the power transmission corridors are divided into regular risk grids, and historical fire data, historical meteorological data, high-precision topographic data, and multi-source vegetation remote sensing data are systematically collected, processed, and integrated. Through machine learning algorithms such as random forest, this module calculates and generates the static flammability baseline level for each risk grid, ultimately forming a "static flammability baseline map" that reflects the inherent fire risk of the region, providing a stable baseline reference for subsequent dynamic risk assessment.

[0165] Real-time multi-source monitoring and data preprocessing module: used for real-time environmental data acquisition and standardization; by deploying multi-sensor terminals such as visible light cameras, infrared thermal imagers and micro-meteorological stations on transmission towers, real-time visual information and micro-meteorological data of the transmission channel are collected simultaneously; through time window alignment and perspective transformation-based spatial coordinate mapping technology, heterogeneous, heterogeneous, and asynchronous monitoring data are precisely spatiotemporally aligned to generate real-time multi-source monitoring data with a unified spatiotemporal benchmark, laying a reliable data foundation for subsequent feature extraction;

[0166] The dynamic risk feature extraction module employs a dual-branch deep learning network architecture to extract key risk features from real-time monitoring data. The visual branch network takes spatiotemporally aligned four-channel images as input, identifies and quantifies smoke, open flame, and high-temperature areas based on the YOLOv5 model, and outputs a visual risk confidence vector. The environmental branch network uses LSTM to process time-series meteorological data, captures trend features that predict the deterioration of environmental flammability, such as sudden increases in wind speed and sudden drops in humidity, and outputs an environmental flammability trend vector, thereby achieving a comprehensive perception of the direct and indirect causes of wildfires.

[0167] The spatiotemporal fusion and risk calculation module is used to intelligently fuse static and dynamic risk information. It has a built-in spatiotemporal attention mechanism, in which the spatial attention unit calculates the spatial weight matrix based on real-time wind direction, slope, and combustible material distribution to quantify the spatial spread tendency of fire. The temporal attention unit dynamically generates temporal fusion weight coefficients based on time periodicity and real-time risk evidence. Through a weighted fusion algorithm, it outputs a comprehensive wildfire risk index (WFRI) for each risk grid that can comprehensively reflect the spatiotemporal risk situation.

[0168] The wildfire spread simulation and graded early warning module is used for risk decision-making. When the WFRI exceeds the threshold, it automatically triggers wildfire spread simulation based on cellular automata and modified Rothermel model to accurately predict the fire path, flame height, and smoke concentration. It also constructs a tripping probability assessment model by combining transmission line parameters (such as conductor-to-ground distance and insulator string length) to dynamically calculate the remaining safe time. Based on the coupling results of WFRI and tripping probability, it initiates graded early warnings from concern to emergency and generates targeted emergency strategies, including line load adjustment and application for shutdown, forming a complete early warning response closed loop.

[0169] The aforementioned power transmission corridor wildfire risk early warning system, based on spatiotemporal feature fusion, forms a closed-loop prevention and control system integrating data perception, feature extraction, intelligent decision-making, and early warning response through collaborative operation. It constructs a precise static risk baseline using multi-dimensional data, providing a stable benchmark for risk assessment. Real-time multi-source monitoring and precise spatiotemporal alignment solve the problem of heterogeneous data fusion. The adopted dual-branch deep learning architecture enables accurate capture of visual fire conditions and environmental trends. The innovative spatiotemporal attention fusion mechanism dynamically optimizes weight allocation, significantly improving the accuracy and adaptability of risk assessment. Based on physical model-based propagation simulation and line tripping probability assessment, it achieves a leap from fire early warning to power grid risk pre-control. It effectively overcomes the data silos and early warning lag problems of traditional monitoring methods, transforming passive response into proactive prevention and control, and significantly improving the intelligence level and emergency response efficiency of power transmission corridor wildfire defense.

[0170] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0171] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion, characterized in that, Includes the following steps: S1. Based on the geographic information of the power transmission channel, the power transmission channel is divided into several risk grids, and combined with historical multi-source data, the static flammability background of each risk grid is calculated to obtain a static flammability background map. S2. Using multi-sensor terminals deployed along the power transmission channel, real-time visual data and real-time micro-meteorological data of the power transmission channel are collected synchronously and spatiotemporally aligned to obtain aligned real-time multi-source monitoring data. S3. Construct a dual-branch spatiotemporal feature extraction network; A dual-branch spatiotemporal feature extraction network was used to extract visual risk features and environmental flammability trend features from aligned real-time multi-source monitoring data, respectively, to obtain visual risk confidence vector and environmental flammability trend vector. S4. Input the static flammability baseline, visual risk confidence vector and environmental flammability trend vector from the static flammability baseline map into the spatiotemporal fusion module, use the spatiotemporal attention mechanism to perform feature fusion, and calculate the comprehensive wildfire risk index. S5. When the comprehensive wildfire risk index exceeds the preset threshold, the wildfire spread simulation model is triggered to predict the wildfire spread trend and the degree of harm to power transmission lines, and graded early warnings and emergency strategies are issued according to the degree of harm.

2. The method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion as described in claim 1, characterized in that, S1 includes the following steps: S11. Establish a coordinate system for the power transmission channel based on the GIS geographic information system. Using the center line of the power transmission line corridor as the reference, extend the preset width to both sides to divide the power transmission channel into several continuous spatial grid units as risk grids. S12. Collect historical fire data, historical meteorological data, topographic data, and vegetation remote sensing data of the power transmission channel; S13. Based on vegetation remote sensing data, invert and calculate the historical combustible load and historical combustible moisture content of each risk grid; combine topographic data and vegetation type to construct a refined vegetation feature map of the power transmission channel. S14. Using the random forest algorithm, with the vegetation fine feature map and historical meteorological data as input, and the fire occurrence probability in historical fire data as the label, a basic fire risk classifier is trained; the trained basic fire risk classifier is used to calculate the static flammability baseline level of each risk grid, and the static flammability baseline level is input into the vegetation fine feature map to form a static flammability baseline map.

3. The method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion as described in claim 2, characterized in that, S13 includes the following steps: S131. Acquire historical multispectral satellite remote sensing images of the power transmission channel area, extract reflectance data in the near-infrared and short-wave infrared bands; collect surface temperature data; S132. Calculate the normalized vegetation index based on reflectance data of near-infrared and short-wave infrared bands. S133. Based on normalized vegetation index and surface temperature data, construct a vegetation water supply index formula; establish a regression model between vegetation water supply index and measured combustible moisture content; establish a regression model between normalized vegetation index and combustible load. S134. Calculate the historical combustible moisture content and combustible load for each risk grid using the above regression model; combine the vegetation type coefficient to obtain the final historical combustible moisture content distribution and historical combustible load distribution for each risk grid. S135. Input the final historical combustible moisture content distribution, historical combustible load distribution, topographic data and vegetation type data into each corresponding risk grid to obtain a refined vegetation feature map.

4. The method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion as described in claim 1, characterized in that, S2 includes the following steps: S21. Deploy multi-sensor terminals on the towers of the power transmission channel; the multi-sensor terminals include visible light cameras, infrared thermal imagers, and micro-weather stations; S22. Real-time video stream data of the power transmission channel is collected using a visible light camera to obtain real-time video stream data; real-time infrared thermal image data is collected using an infrared thermal imager to obtain real-time infrared thermal image data; real-time wind speed, wind direction, temperature, humidity, and air pressure data are collected using a micro-weather station to obtain real-time micro-weather data; the real-time infrared thermal image data and real-time video stream data constitute real-time visual data. S23. Perform spatiotemporal alignment processing on the collected real-time video stream data and real-time infrared thermal image data; map video stream data and micro-meteorological data with different sampling frequencies to the same time window; use the perspective transformation matrix to map the pixel coordinates of the two-dimensional image to the three-dimensional geospatial coordinate system to achieve spatial alignment between visual data and risk grid. S24. Use the spatially aligned real-time video stream data, real-time infrared thermal image data, and real-time micro-meteorological data as aligned real-time multi-source monitoring data.

5. The method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion as described in claim 4, characterized in that, S3 includes the following steps: S31. Construct a dual-branch spatiotemporal feature extraction network; the dual-branch spatiotemporal feature extraction network includes a visual branch network and an environment branch network; the visual branch network adopts a YOLOv5 lightweight convolutional neural network, and adds an infrared channel at the input end of the YOLOv5 lightweight convolutional neural network, superimposing the visible light RGB three channels with the infrared single channel into a 4-channel input; the environment branch network adopts a long short-term memory network; S32. Input the aligned real-time video stream data and real-time infrared thermal image data into the visual branch network, extract visual risk features, and obtain a visual risk confidence vector; the visual risk features include visible light features and infrared features. S33. The aligned micro-meteorological data is input into the environmental branch network. The LSTM network uses its gating mechanism to capture the sudden increase trend of wind speed, the sudden decrease trend of humidity and the continuous increase trend of temperature in the time series, and obtains the environmental flammability trend vector through full-connected layer mapping.

6. The method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion as described in claim 5, characterized in that, S4 includes the following steps: S41. Construct a spatiotemporal fusion module; the spatiotemporal fusion module includes a spatial attention unit and a temporal attention unit; S42. Input the static flammability baseline, visual risk confidence vector, and environmental flammability trend vector into the spatiotemporal fusion module; calculate the spatial weight matrix using the spatial attention unit and calculate the temporal weight coefficient using the temporal attention unit; S43. Based on the spatial weight matrix and time weight coefficient, the static flammability background, visual risk confidence vector and environmental flammability trend vector are weighted and fused to calculate the comprehensive wildfire risk index for each risk grid.

7. The method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion as described in claim 6, characterized in that, S42 includes the following steps: S421. Calculate the ratio of the number of grids located in the current real-time wind direction or at an altitude higher than the current risk grid in all neighboring grids of any risk grid to the total number of neighboring grids of any risk grid, and obtain the wind direction slope weight. S422, The time attention unit is a fully connected neural network; collect historical wildfire case data; the historical wildfire case data includes the time period characteristics, visual risk confidence, environmental flammability trend value and static flammability background of each case; Based on the actual fire spread of each case in the historical wildfire case data, a historical time weighting coefficient is set; the historical time weighting coefficient is set as the label of the historical wildfire case data to obtain labeled historical wildfire case data. The fully connected neural network was trained using labeled historical wildfire case data to obtain the final fully connected neural network. The real-time time period features, visual risk confidence, environmental flammability trend value, and static flammability background are input into the final fully connected neural network to obtain the real-time time weight coefficients.

8. The method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion as described in claim 1, characterized in that, S5 includes the following steps: S51. Set wildfire risk thresholds; monitor the comprehensive wildfire risk index of each risk grid in real time. When the comprehensive wildfire risk index exceeds the wildfire risk threshold, the grid is determined to be the ignition point, triggering the wildfire spread simulation model. S52. Construct a wildfire spread simulation model based on the Rothermel model; introduce a power transmission corridor effect correction coefficient to correct the wind speed parameter in the Rothermel model, and obtain the corrected wildfire spread simulation model. S53. Using the modified wildfire spread simulation model, with the ignition point as the center, combined with real-time micro-meteorological data and topographic data, predict the fire spread path, flame height and smoke concentration distribution within a preset time period in the future. S54. Establish a tripping probability assessment model for transmission lines; calculate the tripping probability of air gap breakdown in transmission lines based on the predicted flame height, smoke concentration, conductor-to-ground distance, and insulator string length. S55. Construct a wildfire risk level; based on the wildfire risk level, combined with the comprehensive wildfire risk index and the probability of power outage, generate corresponding emergency strategies.

9. The method for early warning of wildfire risk in power transmission channels based on spatiotemporal feature fusion as described in claim 8, characterized in that, S54 includes the following steps: S541. Establish a formula for calculating the air gap insulation strength under flame and smoke environments; based on the flame height and smoke concentration distribution, calculate the predicted air breakdown voltage at future moments using the air gap insulation strength calculation formula. S542. Obtain the operating voltage of the transmission line; calculate the tripping probability at each future moment based on the operating voltage of the transmission line and the predicted air breakdown voltage at future moments.

10. A power transmission channel wildfire risk early warning system based on spatiotemporal feature fusion, characterized in that, The system implementing the method for early warning of wildfire risks in power transmission channels based on spatiotemporal feature fusion as described in any one of claims 1-9, comprises: Multidimensional data acquisition and static baseline construction module: Divide the power transmission channel into regular risk grids and collect historical multi-source data; calculate and generate the static flammability baseline level for each risk grid, and construct a static flammability baseline map; Real-time multi-source monitoring and data preprocessing module: By deploying multi-sensor terminals on transmission towers, real-time visual information and micro-meteorological data of the transmission channel are collected synchronously; by time window alignment and perspective transformation-based spatial coordinate mapping technology, aligned real-time multi-source monitoring data is generated. Dynamic risk feature extraction module: Based on aligned real-time multi-source monitoring data, the visual branch network outputs a visual risk confidence vector; the environmental branch network outputs an environmental flammability trend vector by capturing trend features that predict the deterioration of environmental flammability. The spatiotemporal fusion and risk calculation module: The static flammability baseline, visual risk confidence vector, and environmental flammability trend vector from the static flammability baseline map are input into the spatiotemporal fusion module. The spatial attention unit calculates the spatial weight matrix based on real-time wind direction, slope, and combustible material distribution to quantify the spatial spread tendency of the fire. The temporal attention unit generates temporal fusion weight coefficients based on time periodicity and real-time risk evidence. Through a weighted fusion algorithm, the comprehensive wildfire risk index of each risk grid is output. Wildfire spread simulation and graded early warning module: When the comprehensive wildfire risk index exceeds the threshold, the wildfire spread simulation model is triggered to accurately predict the fire path, flame height and smoke concentration; a tripping probability assessment model is constructed in combination with transmission line parameters and the tripping probability is calculated; emergency strategies are generated based on the comprehensive wildfire risk index and the tripping probability.

Images